Ghosts in the machine: AI malware shows why it is time to extend Zero Trust to code

Anthropic's paper reveals that over 80% of the code integrated into their production systems originates from their AI model, Claude. This trend indicates that AI-powered machines can generate, test, adapt, and deploy code at a pace that surpasses human capabilities. Traditional software security measures that rely on human approval and review are becoming obsolete.

When software moves from prompt to execution within minutes, AI-generated code can instantaneously become a script, dependency, automation job, or infrastructure change—activities that typically require human intervention.

AI-assisted analysis can enhance threat detection, but it often produces probabilities rather than definitive policies. In the fast-paced environment created by AI-generated code, "probably suspicious" is insufficient to warrant action. Machines are altering the attack model by automating reconnaissance, vulnerability discovery, exploit generation, payload rewriting, and command sequence adaptation to target environments.

Traditional security measures are built around human constraints and patterns, which do not apply to machine-generated threats.

While software supply chain security has improved, many measures still focus on validating an artifact's properties before execution rather than governing execution itself. Software can meet all established checks and still present risks once executed. An artifact that passes SBOMs, signing, and provenance checks may still violate policies at runtime, while AI-generated scripts could perform their intended functions without triggering alarms. Therefore, a clean dependency list does not guarantee safe behavior.

Post-execution detection is ineffective as it occurs after the risk has entered the environment. By the time suspicious behavior is detected, the software may have already accessed sensitive data, altered system settings, established network connections, or established persistence. AI compacts the window for detection. Instead of asking, "Can we contain this software if it behaves badly?" the focus should shift to, "Should this behavior be permitted to execute in the first place?"

This approach moves the decisive security gate left, akin to the principles of Zero Trust, which rejects implicit trust and requires verification of users, devices, and requests against policy. Before software runs, its expected behavior should be evaluated against security policies. If the behavior aligns with policy, execution can proceed; otherwise, the artifact should be blocked, restricted, isolated, or escalated for further review.

As AI increasingly contributes to both legitimate and malicious code creation, organizations must adopt a Zero Trust approach for code, verifying behavior before execution.